from abc import ABC, abstractmethod import torch import torch.nn as nn from .multimodal_encoder.builder import build_vision_tower from ChatUniVi.constants import * from .cluster import CTM, TCBlock from collections import OrderedDict class MetaModel: def __init__(self, config): super(MetaModel, self).__init__(config) if hasattr(config, "mm_vision_tower"): self.vision_tower = build_vision_tower(config, delay_load=True) self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size) if hasattr(config, "config"): self.use_cluster = config.config["use_cluster"] if self.use_cluster: self.ctm0 = CTM(sample_ratio=config.config["spatial_cluster_rate0"], embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=5) self.block0 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm1 = CTM(sample_ratio=config.config["spatial_cluster_rate1"], embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=3) self.block1 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm2 = CTM(sample_ratio=config.config["spatial_cluster_rate2"], embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=3) self.block2 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm3 = CTM(sample_ratio=config.config["temporal_cluster_rate"], embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=5) self.block3 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) else: self.use_cluster = False def get_vision_tower(self): vision_tower = getattr(self, 'vision_tower', None) if type(vision_tower) is list: vision_tower = vision_tower[0] return vision_tower def initialize_vision_modules(self, model_args, fsdp=None): vision_tower = model_args.vision_tower mm_vision_select_layer = model_args.mm_vision_select_layer mm_vision_select_feature = model_args.mm_vision_select_feature pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter self.config.mm_vision_tower = vision_tower vision_tower = build_vision_tower(model_args) self.config.use_mm_proj = True self.config.mm_hidden_size = vision_tower.hidden_size self.config.mm_vision_select_layer = mm_vision_select_layer self.config.mm_vision_select_feature = mm_vision_select_feature if fsdp is not None and len(fsdp) > 0: self.vision_tower = [vision_tower] else: self.vision_tower = vision_tower if not hasattr(self, 'mm_projector') or not self.mm_projector.weight.size(0): self.mm_projector = nn.Linear(self.config.mm_hidden_size, self.config.hidden_size) if pretrain_mm_mlp_adapter is not None: mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu') def get_w(weights, keyword): return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k} self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector')) def initialize_cluster_modules(self, model_args): self.use_cluster = model_args.use_cluster if self.use_cluster and not hasattr(self, 'ctm0'): self.ctm0 = CTM(sample_ratio=model_args.spatial_cluster_rate0, embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=5) self.block0 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm1 = CTM(sample_ratio=model_args.spatial_cluster_rate1, embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=3) self.block1 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm2 = CTM(sample_ratio=model_args.spatial_cluster_rate2, embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=3) self.block2 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) self.ctm3 = CTM(sample_ratio=model_args.temporal_cluster_rate, embed_dim=self.config.mm_hidden_size, dim_out=self.config.mm_hidden_size, k=5) self.block3 = TCBlock(dim=self.config.mm_hidden_size, num_heads=8) class ChatUniViMetaForCausalLM(ABC): @abstractmethod def get_model(self): pass def get_vision_tower(self): return self.get_model().get_vision_tower() def encode_images(self, images): image_features = self.get_model().get_vision_tower()(images, select_feature="patch") return image_features def positional_encoding(self, x, num_features=1024, max_len=64): p = torch.zeros((1, max_len, num_features)) _x = torch.arange(max_len, dtype=torch.float32).reshape(-1, 1) / torch.pow(10000, torch.arange(0, num_features, 2, dtype=torch.float32) / num_features) p[:, :, 0::2] = torch.sin(_x) p[:, :, 1::2] = torch.cos(_x) x = x + p[:, :x.shape[1], :].to(x.device).to(x.dtype) return x def project(self, image_features, input_type="image"): if self.get_model().use_cluster: if input_type == "image": cluster_image_features = [] token_dict = {'x': image_features, 'token_num': image_features.size(1), 'idx_token': torch.arange(image_features.size(1))[None, :].repeat( image_features.size(0), 1), 'agg_weight': image_features.new_ones(image_features.size(0), image_features.size(1), 1), 'mask': None} token_dict = self.get_model().block0(self.get_model().ctm0(token_dict)) cluster_image_features.append(token_dict["x"]) token_dict = self.get_model().block1(self.get_model().ctm1(token_dict)) cluster_image_features.append(token_dict["x"]) token_dict = self.get_model().block2(self.get_model().ctm2(token_dict)) cluster_image_features.append(token_dict["x"]) image_features = torch.cat(cluster_image_features, dim=1) image_features = image_features.to(self.get_model().mm_projector.weight.dtype) else: cls_features = torch.mean(image_features, dim=1, keepdim=False).unsqueeze(0).clone() token_dict = {'x': cls_features, 'token_num': cls_features.size(1), 'idx_token': torch.arange(cls_features.size(1))[None, :].repeat( cls_features.size(0), 1), 'agg_weight': cls_features.new_ones(cls_features.size(0), cls_features.size(1), 1), 'mask': None} down_dict, token_dict = self.get_model().ctm3(token_dict) events = OrderedDict() max_len = 0 for id, i in enumerate(down_dict["idx_token"][0].tolist()): if i not in events: events[i] = [id] else: events[i].append(id) max_len = len(events[i]) if max_len < len(events[i]) else max_len cluster_image_features = [] token_dict = {'x': image_features, 'token_num': image_features.size(1), 'idx_token': torch.arange(image_features.size(1))[None, :].repeat( image_features.size(0), 1), 'agg_weight': image_features.new_ones(image_features.size(0), image_features.size(1), 1), 'mask': None} token_dict0 = self.get_model().block0(self.get_model().ctm0(token_dict)) token_dict1 = self.get_model().block1(self.get_model().ctm1(token_dict0)) token_dict2 = self.get_model().block2(self.get_model().ctm2(token_dict1)) for id, key in enumerate(events): cur_image_features0 = torch.cat([token_dict0["x"][i] for i in events[key]], dim=0).unsqueeze(0) token_dict = {'x': cur_image_features0, 'token_num': cur_image_features0.size(1), 'idx_token': torch.arange(cur_image_features0.size(1))[None, :].repeat( cur_image_features0.size(0), 1), 'agg_weight': cur_image_features0.new_ones(cur_image_features0.size(0), cur_image_features0.size(1), 1), 'mask': None} cur_token_dict0 = self.get_model().block0(self.get_model().ctm0(token_dict)) cluster_image_features.append(cur_token_dict0["x"]) cur_image_features1 = torch.cat([token_dict1["x"][i] for i in events[key]], dim=0).unsqueeze(0) token_dict = {'x': cur_image_features1, 'token_num': cur_image_features1.size(1), 'idx_token': torch.arange(cur_image_features1.size(1))[None, :].repeat( cur_image_features1.size(0), 1), 'agg_weight': cur_image_features1.new_ones(cur_image_features1.size(0), cur_image_features1.size(1), 1), 'mask': None} cur_token_dict1 = self.get_model().block1(self.get_model().ctm1(token_dict)) cluster_image_features.append(cur_token_dict1["x"]) cur_image_features2 = torch.cat([token_dict2["x"][i] for i in events[key]], dim=0).unsqueeze(0) token_dict = {'x': cur_image_features2, 'token_num': cur_image_features2.size(1), 'idx_token': torch.arange(cur_image_features2.size(1))[None, :].repeat( cur_image_features2.size(0), 1), 'agg_weight': cur_image_features2.new_ones(cur_image_features2.size(0), cur_image_features2.size(1), 1), 'mask': None} cur_token_dict2 = self.get_model().block2(self.get_model().ctm2(token_dict)) cluster_image_features.append(cur_token_dict2["x"]) image_features = torch.cat(cluster_image_features, dim=1) image_features = image_features.to(self.get_model().mm_projector.weight.dtype) else: if input_type == "video": image_features, cls_features = torch.mean(image_features, dim=0, keepdim=False).unsqueeze( 0), torch.mean(image_features, dim=1, keepdim=False).unsqueeze(0) image_features = torch.cat([image_features, cls_features], dim=1) image_features = self.get_model().mm_projector(image_features) return image_features def prepare_inputs_labels_for_multimodal( self, input_ids, attention_mask, past_key_values, labels, images ): vision_tower = self.get_vision_tower() if vision_tower is None or images is None or input_ids.shape[1] == 1: if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1: attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device) return input_ids, attention_mask, past_key_values, None, labels if type(images) is list or images.ndim == 5: concat_images = torch.cat([image for image in images], dim=0) image_features = self.encode_images(concat_images) split_sizes = [image.shape[0] for image in images] image_features = torch.split(image_features, split_sizes, dim=0) image_features = [x.flatten(0, 1) for x in image_features] else: image_features = self.encode_images(images) new_input_embeds = [] new_labels = [] if labels is not None else None cur_image_idx = 0 for batch_idx, cur_input_ids in enumerate(input_ids): if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0: # multimodal LLM, but the current sample is not multimodal cur_input_embeds = self.get_model().embed_tokens(cur_input_ids) cur_input_embeds = cur_input_embeds + ( 0. * self.get_model().mm_projector(vision_tower.dummy_feature)).sum() new_input_embeds.append(cur_input_embeds) if labels is not None: new_labels.append(labels[batch_idx]) cur_image_idx += 1 continue image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0] cur_new_input_embeds = [] if labels is not None: cur_labels = labels[batch_idx] cur_new_labels = [] assert cur_labels.shape == cur_input_ids.shape if len(image_token_indices) > 1: temp = [] cur, pre = image_token_indices[0], image_token_indices[0] for i in image_token_indices: cur = i if cur - pre == 1: temp[-1] = temp[-1] + [cur] else: temp.append([cur]) pre = cur for i in temp: image_token_start = image_token_indices[0] image_token_end = image_token_indices[-1] cur_image_features = [] for _ in i: cur_image_features.append(image_features[cur_image_idx]) cur_image_idx += 1 if len(i) > 2: cur_image_features = torch.stack(cur_image_features, dim=0) cur_image_features = self.project(cur_image_features, input_type="video") t, l, n = cur_image_features.size() cur_image_features = cur_image_features.contiguous().view(t * l, n) else: cur_image_features = torch.stack(cur_image_features, dim=0) cur_image_features = self.project(cur_image_features, input_type="image") t, l, n = cur_image_features.size() cur_image_features = cur_image_features.contiguous().view(t * l, n) if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start - 1]).detach()) cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start - 1:image_token_start])) cur_new_input_embeds.append(cur_image_features) cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_end + 1:image_token_end + 2])) if labels is not None: cur_new_labels.append(cur_labels[:image_token_start]) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype)) cur_new_labels.append(cur_labels[image_token_end:image_token_end + 1]) cur_labels = cur_labels[image_token_end + 2:] else: cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start])) cur_new_input_embeds.append(cur_image_features) if labels is not None: cur_new_labels.append(cur_labels[:image_token_start]) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype)) cur_labels = cur_labels[image_token_end + 1:] if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): cur_input_ids = cur_input_ids[image_token_end + 2:] else: cur_input_ids = cur_input_ids[image_token_end + 1:] elif image_token_indices.numel() > 0: cur_image_features = [] image_token_start = image_token_indices[0] image_token_end = image_token_indices[-1] for _ in image_token_indices: cur_image_features.append(image_features[cur_image_idx]) cur_image_idx += 1 cur_image_features = torch.stack(cur_image_features, dim=0) cur_image_features = self.project(cur_image_features, input_type="image") t, l, n = cur_image_features.size() cur_image_features = cur_image_features.contiguous().view(t * l, n) if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start-1]).detach()) cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_start-1:image_token_start])) cur_new_input_embeds.append(cur_image_features) cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[image_token_end+1:image_token_end+2])) if labels is not None: cur_new_labels.append(cur_labels[:image_token_start]) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype)) cur_new_labels.append(cur_labels[image_token_end:image_token_end+1]) cur_labels = cur_labels[image_token_end+2:] else: cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start])) cur_new_input_embeds.append(cur_image_features) if labels is not None: cur_new_labels.append(cur_labels[:image_token_start]) cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype)) cur_labels = cur_labels[image_token_end+1:] if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): cur_input_ids = cur_input_ids[image_token_end+2:] else: cur_input_ids = cur_input_ids[image_token_end+1:] if cur_input_ids.numel() > 0: if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False): cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids).detach()) else: cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids)) if labels is not None: cur_new_labels.append(cur_labels) cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds] cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0) new_input_embeds.append(cur_new_input_embeds) if labels is not None: cur_new_labels = torch.cat(cur_new_labels, dim=0) new_labels.append(cur_new_labels) if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds): max_len = max(x.shape[0] for x in new_input_embeds) new_input_embeds_align = [] for cur_new_embed in new_input_embeds: cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0) new_input_embeds_align.append(cur_new_embed) new_input_embeds = torch.stack(new_input_embeds_align, dim=0) if labels is not None: new_labels_align = [] _new_labels = new_labels for cur_new_label in new_labels: cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0) new_labels_align.append(cur_new_label) new_labels = torch.stack(new_labels_align, dim=0) if attention_mask is not None: new_attention_mask = [] for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels): new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device) new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device) cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0) new_attention_mask.append(cur_new_attention_mask) attention_mask = torch.stack(new_attention_mask, dim=0) assert attention_mask.shape == new_labels.shape else: new_input_embeds = torch.stack(new_input_embeds, dim=0) if labels is not None: new_labels = torch.stack(new_labels, dim=0) if attention_mask is not None: new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device) attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1) assert attention_mask.shape == new_input_embeds.shape[:2] return None, attention_mask, past_key_values, new_input_embeds, new_labels def initialize_vision_tokenizer(self, model_args, tokenizer): if model_args.mm_use_im_patch_token: tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True) tokenizer.add_tokens([DEFAULT_VIDEO_PATCH_TOKEN], special_tokens=True) self.resize_token_embeddings(len(tokenizer)) if model_args.mm_use_im_start_end: num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN, DEFAULT_VID_START_TOKEN, DEFAULT_VID_END_TOKEN], special_tokens=True) self.resize_token_embeddings(len(tokenizer)) if num_new_tokens > 0: input_embeddings = self.get_input_embeddings().weight.data output_embeddings = self.get_output_embeddings().weight.data input_embeddings_avg = input_embeddings[:-num_new_tokens].mean( dim=0, keepdim=True) output_embeddings_avg = output_embeddings[:-num_new_tokens].mean( dim=0, keepdim=True) input_embeddings[-num_new_tokens:] = input_embeddings_avg output_embeddings[-num_new_tokens:] = output_embeddings_avg if model_args.tune_mm_mlp_adapter: for p in self.get_input_embeddings().parameters(): p.requires_grad = True for p in self.get_output_embeddings().parameters(): p.requires_grad = False if model_args.pretrain_mm_mlp_adapter: mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu') embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight'] assert num_new_tokens == 2 if input_embeddings.shape == embed_tokens_weight.shape: input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:] elif embed_tokens_weight.shape[0] == num_new_tokens: input_embeddings[-num_new_tokens:] = embed_tokens_weight else: raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.") elif model_args.mm_use_im_patch_token: if model_args.tune_mm_mlp_adapter: for p in self.get_input_embeddings().parameters(): p.requires_grad = False for p in self.get_output_embeddings().parameters(): p.requires_grad = False